Stop counting deflected calls

The member has been on hold when the representative picks up, and the work has not started. The question is ordinary — did my MRI get approved, and what will I owe — but the answer is distributed across the adjudication platform, utilization-management portal, benefits accumulator, and CRM. The representative assembles a picture no single screen will show, live, while the member waits.

Repeated across a contact center, that pattern turns skilled service work into systems integration performed by hand.

The AI conversation about that operation is usually framed around deflection rate: the share of calls that never reached a person. Deflection is the wrong unit of account. The consequential event is an operating-model change in which AI resolves routine requests from the member's governed record and specialists own exceptions, escalations, and decisions that require human authority.

The wrong unit of account

Deflection was born in the IVR era and never escaped it. It counts containment, not resolution — a menu tree that frustrates a member into hanging up scores exactly like one that answered the question. Worse, a center that posts a strong deflection number is still shaped the way it always was: the same hiring treadmill, the same five systems on every desk, and a training curriculum whose core competency is toggling between them quickly. The calls got fewer. The organization stayed identical.

That is a procurement outcome, not a transformation.

The retrieval was never the conversation

Watch the eligibility call again and trace where the work actually goes. The representative is doing systems integration by hand, live, with an audience — reconciling multiple systems into one answer while the member waits.

An agent that answers from the record starts where the rep finishes. Before the greeting ends it already holds the claim state, the PA disposition, and where the member stands against the deductible — precisely the things the rep was toggling to find. What is left of the call is the part that was always the point: telling the member, and handling whatever the member says next.

The obvious objection is that members hate AI on the phone, and the objection is earned — two decades of IVR deflection earned it. But look at what those systems actually did. They asked for a member ID the center already had, offered nine menu options that matched nobody's problem, and read scripts written in ignorance of the claim that pended yesterday. Members were not rejecting automation. They were rejecting ignorance with a synthesized voice. An agent that opens already knowing your claim pended, and why, and what happens next, is running the opposite mechanism — which makes the objection to the old one close to evidence for the new one.

Define the transition against your own operation

The design starts with the plan's own call mix, repeat-contact patterns, transfer reasons, after-call work, quality findings, and escalation demand. Routine requests become candidates for governed resolution from the record. Complex, contested, or emotionally sensitive interactions move into a specialist lane with the assembled context, source evidence, and point of escalation already attached.

The measurement model changes with the work. Routine resolution is evaluated for correctness, completeness, repeat contact, and evidence quality. Specialist work is evaluated for resolution authority, exception ownership, transfer quality, and the integrity of the final record. The staffing model follows measured workflow demand instead of a generic deflection target.

What the specialists own

The specialist role centers on contested determinations, active appeals, coordination across providers and discharge plans, grievances, and interactions where empathy and authority matter together. The role carries the power to resolve, longer conversations by design, and a complete operational picture before the handoff begins.

The control surface is equally explicit. Policy defines what the agent may say, which sources it may use, and where it routes the interaction. Authorized reviewers gate consequential workflows, including adverse determinations. The same governed record the agent answers from captures the sources, actions, handoffs, and approvals needed to reconstruct the interaction later.

The procurement question therefore becomes more exact: what operating model will this system produce, which work moves into governed automation, which work moves to specialists, and what evidence proves that the transition is working?